Maximal Domain Independent Representations Improve Transfer Learning
State of the art domain adaptation involves the creation of (1) a domain independent representation (DIRep) trained so that from that representation it is not possible to determine whether the input is from the source domain or the target and (2) a domain dependent representation (DDRep). The original input can then be reconstructed from those two representations. The classifier is trained only on source images using the DIRep. We show that information useful only in the source can be present in the DIRep, weakening the quality of the domain adaptation. To address this shortcoming, we ensure that DDRep is small and thus almost all information is available in the DIRep. We use synthetic data sets to illustrate a specific weakness, which we call the hidden data effect, and show in a simple context how our approach addresses it. We further showcase the performance of our approach against state-of-the-art algorithms using common image datasets. We also highlight the compatibility of our model with pretrained models, extending its applicability and versatility in real-world scenarios.
PDF Abstract
MNIST
SVHN
Office-31
MNIST-M
